Analysis of ChIP-seq Data with mosaics Package

Transcription

1 Analysis of ChIP-seq Data with mosaics Package Dongjun Chung 1, Pei Fen Kuan 2 and Sündüz Keleş 1,3 1 Department of Statistics, University of Wisconsin Madison, WI Department of Biostatistics, University of North Carolina at Chapel Hill Chapel Hill, NC Department of Biostatistics and Medical Informatics, University of Wisconsin Madison, WI September 24, Overview This vignette provides an introduction to the analysis of ChIP-seq data with mosaics package. R package mosaics implements MOSAiCS, a statistical framework for the analysis of ChIP-seq data, proposed in [1]. MOSAiCS stands for MOdel-based one and two Sample Analysis and Inference for ChIP-Seq Data. Based on careful investigation of biases in ChIP-seq data such as mappability and GC content, MOSAiCS has been developed as a flexible mixture modeling approach for detecting peaks of one-sample (ChIP sample) or two-sample (ChIP sample and matched control sample) ChIP-seq data. mosaics package assumes chromosome-wise analysis of ChIP-seq data. In this vignette, we use chromosome 22 of the data from a ChIP-seq experiment of STAT1 binding in interferon-γstimulated HeLa S3 cells from [2] The package can be loaded with the command: R> library("mosaics") 2 Workflow: One-Sample Analysis 2.1 Read Bin-Level Data into the R Environment For the one-sample analysis, MOSAiCS requires four information: bin-level ChIP data, mappability score, GC content score, and sequence ambiguity score. For human genome (HG18) and mouse genome (MM9), we provide mappability score, GC content score, and sequence ambiguity score for each chromosome. Please check for further information about these files and how to download them. Bin-level ChIP data can be obtained from the aligned data (e.g., a file obtained from ELAND) using the perl codes we provide. Users can download these perl codes from Bin-level data obtained from our perl codes can be imported to the R environment with the command: R> bin1 <- readbins(type = c("chip", "M", "GC", "N"), filename = c("./chip_chr22.txt", + "./M_chr22.txt", "./GC_chr22.txt", "./N_chr22.txt")) 1

2 In type, chip, M, GC, N indicate bin-level ChIP data, mappability score, GC content score, and sequence ambiguity score, respectively. User needs to provide the corresponding file names in filename. mosaics package assumes that each file name in filename is provided in the same order as in type. This method also provides information about percentage of bins with ambiguous sequences (i.e., sequence ambiguity score = 1 ) and first/last coordinates of bins before and after preprocessing. R package mosaics provides simple investigation tools for bin-level data. The following command prints out basic information about the bin-level data, such as number of coordinates and total effective tag counts. Total effective tag counts are defined as the sum of tag counts of all bins. This value is usually larger than the number of all the tags used in the analysis. Total effective tag counts provide more useful information about the size of data used in the analysis. R> bin1 Summary: bin-level data (class: BinData) # of coordinates = total effective tag counts = (sum of tag counts of all bins) - input sample is incorporated - mappability is incorporated - GC content is incorporated - uniquely matched tags are assumed print method returns the bin-level data in data frame format. R> print(bin1)[1:10, ] coord tagcount mappability gccontent input plot method provides exploratory plots of mean ChIP tag counts. plottype controls which plot to be drawn. plottype="m" and plottype="gc" generate a plot of mean tag counts versus mappability score and GC content score, respectively. If plottype is not specified, this method plots histogram of mean ChIP tag count. Figure 1, 2, and 3 show examples. One can see that mean tag count increases as mappability score increases. As GC content score increases, mean tag count increases to certain value of GC content score (about 0.6) and then it decreases. R> plot(bin1, plottype = "M") R> plot(bin1, plottype = "GC") R> plot(bin1) 2

3 Mappability vs. ChIP mean tag count ChIP mean tag count mappability Figure 1: Plot of mean ChIP tag counts versus mappability score. 2.2 Fit MOSAiCS We are now ready to fit a MOSAiCS model. The bin-level data above (say, bin1) is used as an input. A MOSAiCS model is fitted with the command: R> fit1 <- mosaicsfit(bin1, analysistype = "OS") analysistype="os" indicates implementation of the one-sample analysis. Without input sample, only the one-sample analysis is permitted. mosaicsfit fits both one-signal-component and two-signal-component models. When calling peaks, user can choose the number of signal components used for the final model. The optimal choice of the number of signal components depends on the characteristics of data. In order to support user in choice of optimal signal model, mosaics package provides Bayesian Information Criterion (BIC) values and Goodness of Fit (GOF) plot of these signal models. The following command shows BIC of signal models, with information about the parameters used in fitting the null (non-enriched) distribution. Lower BIC indicates better model fit and we can see that two-signal-component model has lower BIC in our case. 3

4 GC content vs. ChIP mean tag count ChIP mean tag count GC content Figure 2: Plot of mean ChIP tag counts versus GC content score. R> fit1 Summary: MOSAiCS model fitting (class: MosaicsFit) analysis type: one-sample analysis parameters used: k = 3, meanthres = 0 BIC of one-signal-component model = BIC of two-signal-component model = plot method provides GOF plot. Using the GOF plot, user can compare null model, onesignal-component model, and two-signal-component model, with the actual data. Specifically, user can check which signal model fits the actual data better. Figure 4 shows an example GOF plot. One can see that two-signal-component model provides better fit for our data. R> plot(fit1) 4

5 Histogram of ChIP tag count Frequency e ChIP Tag count Figure 3: Histogram of mean ChIP tag counts. 2.3 Call Peaks [Note: We may be going to include explanation of peak refinement and the procedure to implement it in this section.] We can call peaks with the command: R> peak1 <- mosaicspeak(fit1, signalmodel = "2S", FDR = 0.05, maxgap = 200, + minsize = 50, thres = 10) Info: use two-signal-component model. Info: calculating posterior probabilities... Info: calling peaks... Info: empirical FDR (before thresholding) = 0.05 Info: empirical FDR (after thresholding) = Info: done! Using BIC and GOF plot in the previous section, we can conclude that two-signal-component model fits our data better. signalmodel="2s" indicates two-signal-component model. Similarly, 5

6 Frequency e+05 Actual data Sim:N Sim:N+S1 Sim:N+S1+S Tag count Figure 4: Goodness of Fit (GOF) plot. The plot shows null model (Sim:N), one-signal-component model (Sim:N+S1), and two-signal-component model (Sim:N+S1+S2), with the actual data. one-signal-component model can be specified by signalmodel="1s". Additionally, user can control false discovery rate ( FDR ), distance (in bp) to combine initial peaks ( maxgap ), minimum size (in bp) of the initial peak so that it is claimed as a peak ( minsize ), and minimum ChIP tag count in each bin so that the bin is claimed as a peak ( thres). User needs to choose these paramaters in order to call peaks accurately. We recommend to set maxgap and minsize as the fragment length and bin size, respectively. Initial nearby peaks are merged if the distance between them is less than maxgap. This is because a fragment covers several bins when bin size is chosen to be shorter than the fragment length. The initial peaks of which length are shorter than minsize are removed. This is because very small initial peak (e.g., singleton) is less likely to be a real peak when bin size is chosen to be shorter than the fragment length. When bin size is same as the fragment length, minsize should be smaller than the fragment length. This is because there are many (potentially real) peaks of which size is the same as bin size. thres is employed to filter out initial peaks with small tag counts because these initial peaks might correspond to false discoveries. Optimal choice depends on the sequencing depth of the ChIP-Seq data. If negative value is provided for thres, thresholding is not applied. 6

7 The following command prints out basic information of called peaks such as number of called peaks, median peak length, and empirical false discovery rate. R> peak1 Summary: MOSAiCS peak calling (class: MosaicsPeak) final model: one-sample analysis with two signal components setting: FDR = 0.05, maxgap = 200, minsize = 50, thres = 10 # of peaks = 1690 median peak length = 349 empirical FDR = print method returns the peak calling results in data frame format. This data frame can contain different number of columns, depending on analysistype of mosaicsfit. For example, if analysistype="os", columns are peak start position, peak stop position, peak size, averaged posterior probabilities, minimum posterior probability, averaged ChIP tag counts, and maximum of ChIP tag counts in each peak. Here, posterior probablity of a bin means the probability that the bin is not a peak, conditional on data. Hence, small posterior probability indicates more evidence that the bin is actually a peak. This output is intended as an input for downstream analysis such as motif analysis. R> print(peak1)[1:10, ] peakstart peakstop peaksize avep minp avechipcount e e e e e e e e e e e e e e e e e e e e maxchipcount map GC User can also export peak calling results to text files in diverse file formats. Currently, mosaics package support TXT, BED, and GFF file formats. In the exported file, TXT file format ( type="txt" ) 7

8 includes all the columns that print method returns. type="bed" and type="gff" export peak calling results in standard BED and GFF file format, respectively, where score is summed tag counts in each peak. Peak calling results can be exported in TXT, BED, and GFF file formats, respectively, with the commands: R> export(peak1, type = "txt", fileloc = "./", filename = "OSpeakList.txt", + chrid = "chr22") R> export(peak1, type = "bed", fileloc = "./", filename = "OSpeakList.bed", + chrid = "chr22") R> export(peak1, type = "gff", fileloc = "./", filename = "OSpeakList.gff", + chrid = "chr22") fileloc and filename indicate the directory and the name of the exported file. chrid means chromosome ID. mosaics package assumes chromosome-wise analysis and does not require user to supply chromosome ID during the analysis. In the exported file, the character specified in chrid is used as chromosome ID, which will be the first column in standard BED and GFF file formats. 3 Workflow: Two-Sample Analysis 3.1 Analysis using Mappability and GC Content When input data is additionally provided, the two-sample analysis can be implemented. The procedure to implement the two-sample analysis is analogous to the procedure to implement the onesample analysis. Bin-level data for the two-sample analysis can be imported to the R environment with the command: R> bin2 <- readbins(type = c("chip", "M", "GC", "N", "input"), filename = c("./chip_chr22.txt", + "./M_chr22.txt", "./GC_chr22.txt", "./N_chr22.txt", "./Input_chr22.txt")) In other words, we need one more element, input, in type. The file name of input data also needs to be specified in filename. When input data is provided, plottype="input" generates a plot of mean ChIP tag counts versus input tag count. Moreover, plottype="m input" and plottype="gc input" generate plots of mean ChIP tag counts versus mappability score and GC content score, respectively, conditional on input tag count. R> plot(bin1, plottype = "input") R> plot(bin1, plottype = "M input") R> plot(bin1, plottype = "GC input") In order to fit MOSAiCS model for the two-sample analysis, when calling mosaicsfit method, user should specify analysistype="ts" instead of analysistype="os". R> fit2 <- mosaicsfit(bin2, analysistype = "TS") Peak calling can be done exactly in the same way as before. Also all the other methods can still be used as in the one-sample analysis case. R> peak2 <- mosaicspeak(fit2, signalmodel = "2S", FDR = 0.05, maxgap = 200, + minsize = 50, thres = 10) 8

9 3.2 Analysis without using Mappability and GC Content Application of MOSAiCS methods to the real data showed that consideration of mappability and GC content in the model improves sensitivity and specificity of peak calling [1]. However, mosaics package still permits the two-sample analysis without considering mappability and GC content. In order to fit MOSAiCS model for the two-sample analysis without considering mappability and GC content, analysistype="io" needs to be specified when calling mosaicsfit method. R> fit3 <- mosaicsfit(bin2, analysistype = "IO") Furthermore, in this case, mappability score, GC content score, and sequence ambiguity score might not be incorporated when importing bin-level data. User can import bin-level data and fit MOSAiCS model for the two-sample analysis without mappabilibity and GC content, with the commands: R> bin4 <- readbins(type = c("chip", "input"), filename = c("./chip_chr22.txt", + "./Input_chr22.txt")) R> fit4 <- mosaicsfit(bin4, analysistype = "IO") Note that with these commands mosaics package skips initial processing of bin-level data using mappability score, GC content score, and sequence ambiguity score. We recommend to incorporate mappability score, GC content score, and sequence ambiguity score as long as they are available to users. 4 Tuning of MOSAiCS Parameters For the two-sample analysis, mosaics package permits user to tune the parameters used in fitting the null distribution of MOSAiCS model. When both mappability score and GC content score are available (i.e., the two-sample analysis using mappability and GC content), user can control two tuning parameters, s and d. s means the cut-off to determine lower and higher values of input data. d means the order used in the transformation of input data. Please see [1] for further details about model parameters. For the two-sample analysis using mappability and GC content, mosaics package uses the following parameter setting as default: R> fit3_ts <- mosaicsfit(bin2, analysistype = "TS", s = 2, d = 0.25) When mappability score and GC content score are not available (i.e., the two-sample analysis without using mappability and GC content), user can tune only one parameter, d. For the twosample analysis without using mappability and GC content, mosaics package uses the following parameter setting as default: R> fit3_io <- mosaicsfit(bin2, analysistype = "IO", d = 0.25) 5 Conclusion and Ongoing Development R package mosaics provides effective tools to read and investigate ChIP-seq data, fit MOSAiCS, and claim peaks. We are continuously working on improving mosaics package further in more friendly user interface and more effective and diverse investigation tools. 9

10 References [1] Kuan, PF, D Chung, JA Thomson, R Stewart, and S Keleş (2010), A Statistical Framework for the Analysis of ChIP-Seq Data, submitted ( 19/). [2] Rozowsky, J, G Euskirchen, R Auerbach, D Zhang, T Gibson, R Bjornson, N Carriero, M Snyder, and M Gerstein (2009), PeakSeq enables systematic scoring of ChIP-Seq experiments relative to controls, Nature Biotechnology, 27,